⚡ electrical engineering

Neural Configuration-Space Barriers for Manipulation Planning and Control

This paper proposes a unified motion planning and control framework for high-dimensional robot manipulators that utilizes learned neural configuration-space distance functions as distributionally robust barriers to ensure safety and efficiency in cluttered, dynamic environments using only onboard sensor data.

Original authors: Kehan Long, Ki Myung Brian Lee, Nikola Raicevic, Niyas Attasseri, Melvin Leok, Nikolay Atanasov

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kehan Long, Ki Myung Brian Lee, Nikola Raicevic, Niyas Attasseri, Melvin Leok, Nikolay Atanasov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guide a long, flexible snake (a robot arm) through a crowded room filled with moving furniture. Your goal is to get the snake's head to a specific spot without it bumping into anything. This is the challenge this paper tackles, but for high-tech robot arms in places like warehouses.

Here is the simple breakdown of their solution, using everyday analogies:

1. The Problem: The "Blindfolded" Planner

Traditional robot planners work like a person trying to walk through a dark room by taking tiny, cautious steps. They constantly stop to check, "Is there a wall here? Is there a chair there?"

The Issue: In a room full of moving obstacles, checking every single step takes too long. By the time the robot finishes checking, the furniture might have moved, or the robot might have already crashed. It's like trying to drive a car while constantly stopping to check the map for every inch of road.

2. The Innovation: The "Safety Bubble"

Instead of checking every tiny step, the authors teach the robot a new trick: Configuration-Space Distance Functions (CDFs).

The Analogy: Imagine the robot arm is a person, and the "Configuration Space" is a map of every possible pose that person could make.
The Magic: The robot learns to draw invisible "Safety Bubbles" around itself on this map. Inside a bubble, the robot knows for a fact it won't hit anything.
The Benefit: Instead of checking every single step, the robot just jumps from one bubble to the next. It's like hopping across stepping stones in a river instead of wading through the water. This makes planning incredibly fast because the robot doesn't need to check for collisions as often.

3. The Catch: The "Imperfect Crystal Ball"

There's a problem with these bubbles: The robot learns them using a neural network (a type of AI) and sensors (cameras).

The Issue: AI makes mistakes, and cameras get noisy. The robot might think a bubble is safe when it's actually too close to a wall. If the robot trusts the bubble blindly, it might crash.
The Real World: The obstacles (furniture) are also moving, and the robot's view of them isn't perfect.

4. The Solution: The "Paranoid Safety Filter"

To fix the uncertainty, the authors created a Distributionally Robust Control Barrier Function (DR-CBF).

The Analogy: Imagine the robot has a "Safety Filter" that acts like a paranoid bodyguard.
- The bodyguard knows the AI's map might be slightly wrong.
- The bodyguard knows the camera might be blurry.
- The bodyguard knows the furniture might move faster than expected.
How it works: When the robot tries to move, the bodyguard doesn't just look at the "perfect" map. It looks at a "worst-case scenario" version of the map (within a reasonable range of error). If there's any chance the bubble is too small or the obstacle is closer than it looks, the bodyguard says, "Stop! Slow down! Move away!"
The Result: The robot might take a slightly less direct path or move a bit more cautiously, but it guarantees it won't crash, even if its sensors are noisy or the environment is chaotic.

5. The Proof: Real-World Testing

The team tested this on a real robot arm (a UFactory xArm6) in a cluttered room with moving obstacles.

Planning: The "Bubble" method was much faster than traditional methods because it took huge, safe leaps instead of tiny, cautious steps.
Control: When obstacles moved toward the robot, the "Paranoid Bodyguard" filter made the robot dodge defensively and then smoothly return to its path, successfully reaching the goal without crashing. Traditional methods often failed in these dynamic situations.

Summary

This paper gives robots a super-fast map (the bubbles) to plan their route and a super-cautious safety guard (the robust filter) to handle the fact that the map and the sensors aren't perfect. The result is a robot that can move quickly and safely through messy, changing environments without needing to stop and think about every single millimeter.

Technical Summary: Neural Configuration-Space Barriers for Manipulation Planning and Control

Problem Statement

High-degree-of-freedom (DoF) robotic manipulators operating in cluttered, dynamic environments face a dual challenge: ensuring computational efficiency in motion planning while maintaining robust safety guarantees during execution. Traditional sampling-based planners (e.g., RRT, PRM) rely on frequent, expensive collision checks, creating a bottleneck in high-dimensional configuration spaces. Furthermore, in dynamic settings with sensor noise and unpredictable obstacles, planning alone is insufficient; real-time control strategies must account for modeling errors and observation uncertainties to prevent collisions. Existing methods often struggle to balance the need for precise, differentiable safety representations with the computational constraints of real-time operation.

Methodology

The authors propose a unified framework integrating motion planning and control using Neural Configuration-Space Distance Functions (CDFs). The approach consists of three primary components:

1. Neural CDF Representation

The method learns two distinct neural network models to represent safety in the configuration space ( $Q$ ):

Environment CDF ( $f_c$ ): Estimates the minimum joint-space distance from a configuration $q$ to a workspace obstacle point $p$ .
Self-Collision CDF ( $f_{sc}$ ): Estimates the minimum joint-space distance from $q$ to the set of self-colliding configurations.
These functions are trained as multi-layer perceptrons (MLPs) to provide fast, differentiable distance and gradient queries, approximating the local free configuration space.

2. Bubble-CDF Planner

To improve planning efficiency, the authors introduce a Rapidly-exploring Bubble Graph (RBG) algorithm. Instead of treating nodes as single configurations, the planner constructs "bubbles" (spherical regions) in the configuration space.

Bubble Construction: The radius of a bubble centered at a safe configuration $q$ is dynamically determined by the neural CDF barrier value: $r(q) = \hat{h}(q, t) - \eta$ , where $\eta$ is a safety margin.
Efficiency: Any configuration within a bubble is guaranteed to be collision-free (subject to the barrier approximation). This allows the planner to explore the space by connecting overlapping bubbles, drastically reducing the number of explicit collision checks required compared to edge-based RRT.
Trajectory Optimization: Once a path of bubbles is found, a continuous trajectory is generated by optimizing a sequence of Bézier curves. The control points of these curves are constrained to lie within the respective bubbles, ensuring the entire trajectory remains collision-free via the convex hull property, formulated as a convex Quadratically Constrained Quadratic Program (QCQP).

3. Distributionally Robust Control Barrier Function (DR-CBF)

For real-time execution, the system employs a Distributionally Robust Control Barrier Function (DR-CBF) to track the planned trajectory while ensuring safety under uncertainty.

Uncertainty Modeling: The approach acknowledges uncertainties arising from neural network approximation errors, point-cloud sensor noise, and dynamic obstacle motion.
Wasserstein Ambiguity Set: Instead of assuming a known probability distribution for these errors, the method constructs a Wasserstein-1 ambiguity set around the empirical distribution of uncertainty samples (generated via Monte Carlo dropout and point-cloud velocity estimation).
Safety Filter: The controller solves a Quadratic Program (QP) online that minimizes the deviation from a nominal PD controller while satisfying a distributionally robust safety constraint. This ensures that the probability of violating the safety set remains below a user-specified risk tolerance ( $\epsilon$ ), even under worst-case distributions within the ambiguity set.

Key Contributions

Unified Neural CDF Barrier: A learned representation that estimates distances to both external obstacles and self-collisions, enabling fast distance and gradient queries from point-cloud observations.
Distributionally Robust Control: A novel DR-CBF formulation for manipulator control that explicitly accounts for modeling errors, sensor noise, and dynamic obstacles without requiring explicit knowledge of the underlying uncertainty distribution.
Bubble-Based Planning: A sampling-based planner that utilizes neural CDF queries to construct a graph of safe configuration-space bubbles, significantly reducing collision-checking overhead in high-dimensional spaces.
Experimental Validation: Comprehensive evaluation on a 6-DoF UFactory xArm6 manipulator in both simulation and real-world dynamic environments, demonstrating the integration of the planner and controller.

Results

The paper presents extensive evaluations comparing the proposed method against baselines including RRT variants (CDF-RRT, SDF-RRT, Lazy-RRT), MPPI, and standard CBF controllers.

Planning Efficiency: In both 2-DoF and 6-DoF simulations, the Bubble-CDF planner reduced the number of collision checks by an order of magnitude (e.g., ~154 checks vs. ~1,759 for Lazy-RRT in 2-DoF; ~279 vs. ~2,741 in 6-DoF) while maintaining comparable path lengths. It also achieved the lowest overall runtime, including post-processing.
Control Robustness: In dynamic environments with moving obstacles:
- Standard PD controllers failed frequently (success rates ~11-18%).
- Standard CBF-QP controllers showed improved performance but dropped significantly in dynamic settings (success rates ~64-70%) due to unmodeled uncertainties.
- The proposed PD + DR-CBF-QP achieved 100% success in both static and dynamic simulations and real-world experiments.
Real-World Performance: Experiments on the physical xArm6 confirmed the simulation results, showing the robot successfully navigating cluttered environments and executing defensive maneuvers to avoid dynamic obstacles while tracking the planned trajectory.

Significance and Claims

The paper claims that this approach provides a practical solution for robotic manipulation in environments where accurate models are unavailable and obstacles move unpredictably. By integrating learned CDFs into both planning and control, the method achieves:

Computational Efficiency: Substantial reduction in collision-checking operations through the use of configuration-space bubbles.
Robust Safety: The ability to maintain safety guarantees despite neural network approximation errors and sensor noise, achieved through distributionally robust optimization rather than conservative offline margins.
Real-Time Applicability: The framework operates effectively on onboard sensors (point clouds) and supports real-time control rates (50 Hz), making it suitable for industrial settings like warehouses and assembly lines where dynamic replanning is necessary.

The authors emphasize that their method does not rely on detailed environmental models and can be extended to other manipulators operating in dynamic environments, aiming to improve productivity and safety in industrial applications.